Understanding the Speedster7t clock architecture is an important part of planning FPGA designs that combine high-speed fabric logic with interfaces such as Ethernet, PCIe, GDDR6, DDR4, and the network on chip (NoC). The architecture provides several ways to generate and distribute clocks, but each network has a defined reach and resource limit. Choosing an appropriate path helps keep clocking aligned with where logic resides and how it connects to interface subsystems.
Speedster7t fabric contains reconfigurable logic blocks (RLBs) and machine learning processors (MLPs), arranged in columns and grouped into clock regions. Its clock system is divided into four broad areas: clock generation, the global core network, interface clocks, and dedicated clocks for subsystems such as the NoC and memory interfaces.
Clock generation: oscillators and PLLs
The device includes 16 on-chip oscillators, located in its four corners. They provide free-running 2 GHz and 100 MHz clocks with statically configured dividers for common ratios. The oscillators are not currently configurable in ACE.
For user-configured clock generation, Speedster7t provides 16 general-purpose fractional-N PLLs, four per corner. A PLL can produce up to four output clocks, with reference clocks available from dedicated clock I/O, other PLLs, and clock sources elsewhere on the device. Outputs can be routed to the global fabric network, nearby GPIO, interface subsystems, or another PLL. Cascading PLLs can support clock distribution across the device.
PLL setup is managed through the ACE I/O Designer Toolkit. Designers select the reference input, PLL location, output frequencies, and which outputs are exposed to the FPGA fabric. If an output is intended only for an interface subsystem, it can be left unexposed to the fabric clock network. This distinction matters because fabric use consumes clock-network resources.
Dedicated PLLs serve specific interfaces. The NoC has dedicated PLLs that generate its required clock from a 200 MHz reference. GDDR6 and DDR4 PHYs also have built-in PLLs for interface-specific clocking, while SerDes quads use PLLs to derive receive clocks. These dedicated PLLs are not interchangeable with general-purpose fabric clocks; their availability and routing depend on the associated subsystem.
How the global core clock network distributes clocks
The global core network is a balanced, low-skew H-tree intended to reach the FPGA fabric. At the top and bottom of the device, Clock Hub 0 (H0) collects clock inputs and passes selected signals toward the central trunk. Clock Hub 1 (H1), at the center, combines inputs from the two H0 hubs and can also accept signals from fabric data interconnect. H1 generates 48 global clocks for distribution to clock regions on both sides of the core.
At each half-row of clock regions, Clock Hub 2 (H2) selects from global clocks and certain data signals, then supplies a 16-clock regional bus. H2 includes optional clock division, dynamic gating, and glitchless switching between clock sources. Its dividers offer static divide-by-2, 4, 6, or 8 settings, or can be bypassed. These features can help adapt a clock for a region, but they do not remove the need to account for the finite number of clock tracks.
Clock Hub 3 (H3), positioned within each clock region, distributes selected clocks through that region and toward adjacent regions. In practical terms, the global network is the option to consider when clocked logic must span regions; regional and local choices are more appropriate when the endpoints are geographically limited.
Interface clocks have a narrower reach
Interface clocks enter the fabric from its edges to support logic associated with nearby interface subsystems. Clocks entering from the north or south drive mini-trunks; those entering from the east or west drive branch clocks. Unlike the global network, an individual interface clock cannot reach every part of the device or cross the central trunk and north-south delimiter.
For mini-trunk clocks, the guide recommends keeping the related fabric logic within the cluster where the clock enters. Branch clocks are intended for nearby logic and are limited to the clock region where they enter. This makes interface clocks useful for keeping fabric logic in the same clock domain as an adjacent interface, but unsuitable as a general replacement for the global core network.
Route fabric-generated signals deliberately
ACE handles much of the clock selection and routing, but designers can use the PDC set_clock_type constraint to guide how a data-generated clock or data signal is routed. The appropriate option depends on endpoint location, not simply signal fan-out:
- -data_center: Routes through H1 for endpoints in multiple clock regions. It can provide balanced distribution across the core, but uses central clock resources and can add insertion delay.
- -data_region: Routes through the regional H2 for endpoints within one clock region. This confines the clock-resource impact to that region.
- -data_local: Intended for very low-fan-out cases, such as a derived clock feeding a small group of nearby clock pins. It can have lower insertion delay, but may produce greater skew as fan-out and endpoint separation increase.
Data signals such as resets and enables are generally routed on the data interconnect. When a high-fan-out signal is placed on the clock network, it can reach multiple endpoints with balanced timing, but consumes clocking capacity. The architecture also has limits on the number of clocks, resets, and enables available at logic cells, RLBs, and within clock regions. In particular, routing resets, enables, or data signals through clock resources reduces capacity for unique clocks. Designs with multiple derived clocks should consider the built-in dividers, gates, and switches where appropriate.
After an ACE compilation, routed-clock reports provide information about clock relationships, constraints, clock regions, and how clocks are distributed. Reviewing these reports helps confirm whether the implemented routes match the intended clock plan.
Plan reset distribution alongside clocking
Speedster7t provides a global reset from the FPGA configuration unit (FCU), external reset inputs through clock I/O, and resets generated in fabric logic. The FCU reset holds the fabric in reset during configuration; clock gating is released glitchlessly as the device transitions toward user operation. Subsystems synchronize reset de-assertion so that release is synchronous within each subsystem.
Reset routing also has physical considerations. A reset from a clock I/O bank can serve multiple interface subsystems, while a locally generated reset is supplied separately to each adjoining subsystem. To distribute one fabric-generated reset to several interfaces, the design must replicate and route the signal to each destination. GPIO is a special case: its I/O flops can use a reset from a clock I/O bank or a local AUX pad, and an AUX-pad reset is limited to that GPIO bank.
Practical implications for FPGA design
Successful clock planning starts by matching the network to the location and reach of its endpoints. Use global clocks for broad fabric distribution, interface clocks for nearby interface logic, and regional or local routing for signals with constrained destinations. Treat clock tracks as a limited design resource, especially when resets, enables, or fabric-derived clocks share the network. Finally, verify the routed implementation in ACE reports and plan reset behavior as part of the same architecture exercise. These choices give engineers a clearer basis for balancing reach, insertion delay, skew, and clock-resource availability in Speedster7t designs.